https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119474
--- Comment #10 from rguenther at suse dot de <rguenther at suse dot de> --- On Mon, 31 Mar 2025, ams at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119474 > > --- Comment #9 from Andrew Stubbs <ams at gcc dot gnu.org> --- > This patch fixes the -O1 failure, for *this* testcase: > > diff --git a/gcc/tree.cc b/gcc/tree.cc > index eccfcc89da40..4bfdb7a938e7 100644 > --- a/gcc/tree.cc > +++ b/gcc/tree.cc > @@ -7085,11 +7085,8 @@ build_pointer_type_for_mode (tree to_type, machine_mode > mode, > if (to_type == error_mark_node) > return error_mark_node; > > - if (mode == VOIDmode) > - { > - addr_space_t as = TYPE_ADDR_SPACE (to_type); > - mode = targetm.addr_space.pointer_mode (as); > - } > + addr_space_t as = TYPE_ADDR_SPACE (to_type); > + mode = targetm.addr_space.pointer_mode (as); > > /* If the pointed-to type has the may_alias attribute set, force > a TYPE_REF_CAN_ALIAS_ALL pointer to be generated. */ > @@ -7121,6 +7118,7 @@ build_pointer_type_for_mode (tree to_type, machine_mode > mode, > TYPE_REF_CAN_ALIAS_ALL (t) = can_alias_all; > TYPE_NEXT_PTR_TO (t) = TYPE_POINTER_TO (to_type); > TYPE_POINTER_TO (to_type) = t; > + TYPE_ADDR_SPACE (t) = TYPE_ADDR_SPACE (to_type); > > /* During LTO we do not set TYPE_CANONICAL of pointers and references. */ > if (TYPE_STRUCTURAL_EQUALITY_P (to_type) || in_lto_p) > > > This is clearly wrong though, because the address space that the pointer is > *in* doesn't have to be the same as the one it *points to*, so I need a better > solution. > > Any suggestions where to start? Well, you have to fixup the oacc lowering process to more "properly" move objects from one address-space to another. Your earlier patch was a good first step in principle.
